Cardiac arrhythmias present as changes in cardiac electrical activity that is measured by electrocardiograms (ECGs). Well-designed deep Convolutional Neural Networks (CNNs) can detect heart abnormalities from ECGs without relying on manual feature engineering methods. In this paper, we automate the process of CNN structure design and present two CNN models for cardiac abnormality detection by using Neural Architecture Search (NAS) techniques. These techniques perform a very efficient search over an extremely large search space of operations in CNNs. Our ECUTeam proposes a CNN architecture designed through the well-known architecture search method, Differentiable Architecture Search (DARTS), and achieved the challenge metric of 0.65, 0.63, 0.64, 0.59 on 12 lead, 6 lead, 3 lead, and 2 lead ECGs, respectively, using 5-fold cross-validation on the training dataset. The second neural model constructed by implementing an improved version of DARTS, Robust DARTS (R-DARTS), on 1-dimensional ECGs performed slightly better and reported the challenge metric of 0.68, 0.64, 0.64, 0.62 for 12, 6,3,2 lead ECG signals employing 5-fold cross-validation on the training dataset. Our proposed neural models are made by a stack of 20 cells, each cell containing convolution and pooling layers, constructed and optimized by implementing the NAS search strategies.